Speech Recognition Accuracy via User Profile Context Integration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face challenges in accurately recognizing rare or personalized words, such as contact names and device names, due to their reliance on generic training data, which can lead to errors in understanding user-specific inputs.
Innovation Solution
The implementation of a neural-based approach that leverages contextual information from user profiles, utilizing a multi-head attention mechanism to enhance the relevance of context data in speech recognition, thereby improving accuracy and reducing the need for external fusion language models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If generic training data is used for speech recognition, then the system can operate with simple architecture and low computational requirements, but the accuracy in recognizing rare or personalized words deteriorates
Solution Approach 1:
The system performs preliminary actions by integrating user profile data (contact names, device names, personalized information) into the speech recognition process before actual recognition occurs. This allows the model to have prior knowledge of rare and personalized words, improving recognition accuracy without requiring complex external fusion models during runtime.
Solution Approach 2:
The patent merges generic training data with user-specific context data into a unified speech recognition model. By combining these data sources during training, the system achieves both general speech recognition capability and specialized recognition of rare/personalized words, resolving the contradiction between accuracy and complexity.
2Measurement precision
If external fusion language models are used to improve recognition of rare words, then speech recognition accuracy improves, but memory footprint and computational resources increase
Solution Approach 1:
The patent combines user profile data directly into the speech recognition model training process, merging general speech data with personalized context. This integration eliminates the need for separate external fusion language models, thereby improving recognition accuracy while maintaining efficient memory usage.
Solution Approach 2:
The system extracts and utilizes only the necessary user-specific context information (contact names, device names, personalized terms) from user profiles, integrating selectively rather than processing entire external models. This extraction approach improves accuracy for rare words without the heavy memory burden of complete external fusion models.
Data Source
AI summary
Techniques for performing automatic speech recognition (ASR) are described. In some embodiments, an ASR component integrates contextual information from user profile data into audio encoding data to predict a token(s) corresponding to a spoken input. The user profile data may include personalized words, such as, contact names, device names, etc. The ASR component determines word embedding data using the personalized words. The ASR component is configured to apply attention to audio frames that are relevant to the personalized words based on processing the audio encoding data and the word embedding data.